Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Not a member yet
1071 research outputs found
Sort by
Hybrid Video Transcription Summarization with a BERT-Based Clustering and BART
No full textThe use of video as a medium for information and education is rapidly increasing across online platforms. However, long durations and unstructured delivery often hinder audiences from grasping the core message, presenting challenges for the development of automatic summarization methods for monologues, interviews, and podcasts. Extractive methods often yield less coherent summaries, while abstractive methods may overlook important details. To address this issue, this study proposes a hybrid approach combining extractive and abstractive techniques. In the extractive stage, sentences are represented using BERT embeddings and clustered using two methods, namely K-Means Clustering and Hierarchical Clustering (agglomerative). The abstractive stage then employs the BART model to generate summaries that are more coherent and informative. Experimental evaluations on 20 Human Metapneumovirus (HMPV) videos indicate the strongest performance on monologues, with ROUGE-1 of 57%, ROUGE-2 of 30%, and ROUGE-L of 32%. Although lower performance was observed for interviews and podcasts due to dynamic interactions and frequent speaker shifts, the hybrid approach consistently surpassed extractive-only and abstractive-only baselines. These results highlight the effectiveness of the hybrid approach and its potential for developing more adaptive video summarization in the future
Improving Vehicle Payment Method Classification Using XGBoost with SMOTE and SHAP Interpretation
No full textClass imbalance in vehicle payment method classification can cause predictive models to become biased toward the majority. This study aims to build a classification model for automotive consumer payment methods using Extreme Gradient Boosting (XGBoost), with class balancing handled through the Synthetic Minority Over-sampling Technique (SMOTE) and Adaptive Synthetic Sampling (ADASYN), and model interpretability performed using SHAP (SHapley Additive Explanations). The dataset consisted of 11,011 records and 13 attributes derived from Toyota vehicle delivery order transactions. Results show that the XGBoost model without balancing achieved 67.37% accuracy but only 0.24 recall for the Cash class. After applying SMOTE, the recall for the Cash class improved to 0.58, while ADASYN produced a similar improvement at 0.59, with overall accuracy maintained at around 61–62% and a stable ROC-AUC of 0.65. Feature importance and SHAP analysis identified c_vehicle_model and c_city as the most influential factors in predicting the payment method. From a business perspective, the improved ability to detect cash customers reduces the risk of misclassification and enables dealers to better segment customer payment preference. This supports more effective marketing campaigns, sales strategies, and financing risk management. The combination of XGBoost, SMOTE, ADASYN, and SHAP has proven effective in handling imbalanced data while offering transparent interpretability of predictions, making it a practical foundation for data-driven decision-making in the automotive industry
Optimization of a New Adaptive Stacking Ensemble Model Integrated with IoT for Stress Level Detection Based on Physiological Signals
No full textMental health issues among college students are receiving increasing attention, particularly because of academic and social pressures and the impact of technology use. This study aims to develop a real-time stress level prediction model using a New Adaptive Stacking Ensemble approach based on physiological data and IoT devices. The data included heart rate, SpO₂, body temperature, and systolic and diastolic blood pressure. Five machine learning algorithms are used as base models: SVM, C4.5, Decision Tree, KNN, and Random Forest. The MLP serves as the meta-model, which is then optimized using Optuna. The model training process begins with pre-processing, feature standardization using StandardScaler, and data balancing using SMOTE. The results showed that the stacking model with the MLP meta-model achieved an accuracy of 90.00% under the individual Random Forest and KNN models, and increased to 97.00% after hyperparameter optimization. This model was then integrated with IoT devices using MAX30102, MLX90614, and digital tensiometer sensors, as well as a Streamlit interface to display real-time stress classification results. The system built not only excels in accuracy but can also be implemented to directly detect stress levels, thereby potentially supporting early intervention and mental health promotion in campus environments
Plant Disease Identification Using Image Processing: A Systematic Literature Review
No full textThis article is a literature review focusing on plant disease identification using image processing techniques. This review aims to provide a comprehensive analysis of dataset sources, preprocessing methodologies, segmentation techniques, feature extraction processes, and various classification methods, along with their associated accuracies. It also discusses challenges encountered and potential future research directions. Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) protocol, a literature search was conducted in the Scopus database to obtain primary studies. The search covered Scopus-indexed journals and proceedings published by IEEE, Elsevier, Springer, MDPI, and ACM between 2019 and 2025. The initial identification phase yielded 9,286 studies screened. Further screening was performed based on specific eligibility criteria, including relevance to the topic, year of publication, subject area, document type, and articles written in English, resulting in the selection of 82 studies for the review. The findings indicate that the most commonly used dataset is PlantVillage, followed by field data. The dominant preprocessing techniques include image enhancement and augmentation. For segmentation and feature extraction, the most frequently used methods were k-means and CNN, respectively. Sixty-one studies achieved an accuracy exceeding 90%. However, several key challenges remain: data limitations, methodological issues, and practical constraints. Future research should focus on developing more representative datasets, hybrid approaches that integrate classical and deep learning methods, and lightweight, adaptive decision support systems suitable for real-world agricultural applications. This review supports continued progress in this field by providing valuable insights for researchers developing image-based methods for identifying plant diseases
Multi-Process Data Mining with Clustering and Support Vector Machine for Corporate Recruitment
No full textHaving an efficient and accurate recruitment process is very important for a company to attract candidates with professionalism, a high level of loyalty, and motivation. However, the current selection method often faces problems due to the subjectivity of assessing prospective employees and the long process of deciding on the best candidate. Therefore, this research aims to optimize the recruitment process by applying data mining techniques to improve efficiency and accuracy in candidate selection. The method used in this research utilizes a multi-process Data Mining approach, which is a combination of clustering and classification algorithms sequentially. In the initial stage, the K-Means algorithm is applied to cluster candidates based on administrative selection data, such as document completeness and reference support. Next, a classification model was built using a Support Vector Machine (SVM) to categorize the best candidates based on the results of psychological tests, medical tests, and interviews. The experimental results show that the SVM model produces high evaluation scores, with an AUC of 87%, Classification Accuracy (CA) of 90%, F1-score of 89%, Precision of 91%, and Recall of 90%. With these results, it can be concluded that this model is able to improve accuracy in the employee selection process and help companies make more measurable and data-based recruitment decisions
A New Triple-Weighted K-Nearest Neighbor Algorithm for Tomato Maturity Classification
No full textAs climatic products, tomatoes are highly sensitive to harvesting and processing. The sorting of tomatoes can be significantly improved by utilizing Hue Saturation Value (HSV) color features that are classified using neighboring algorithms, such as K-Nearest Neighbor (KNN), Weighted K-Nearest Neighbor (W-KNN), and DW-KNN. However, the DW-KNN algorithm does not consider the relative relationship between the farthest, nearest, and surrounding neighbors, which may impact the classification accuracy, particularly in datasets with uneven distributions. This study proposes a Triple Weighted K-Nearest Neighbor (TW-KNN) algorithm for tomato image classification. This algorithm effectively handles the problem of sensitivity and outliers in the data distribution and considers the relationship between neighboring distances. The classification data consisted of 400 tomato images with five maturity levels divided into training and testing sets using k-fold cross-validation. Tests were conducted using several variations of parameter k, namely 4, 6, 9, and 15, to evaluate the classification performance. The results show that the proposed TW-KNN algorithm consistently outperforms other methods by producing better classification results. This is demonstrated by an accuracy rate of 95.52% across different values of k. The superior performance of the TW-KNN highlights its ability to provide robust and stable classification results compared to conventional KNN variants. This finding indicates that the TW-KNN is more effective in consistently classifying tomato fruits, making it a promising approach for automated fruit sorting applications
Hyperparameter Tuning with Optuna to optimize the YOLOv11n Model for Weed Detection
No full textAccurate weed detection is essential for maintaining the cleanliness and aesthetic appeal of residential yards. This study aimed to optimize YOLOv11n, a lightweight object detection model, to achieve high precision in weed identification under real-world conditions. The novelty of this study lies in the application of Optuna, an automatic hyperparameter optimization framework, to enhance model performance while maintaining computational efficiency for resource-limited devices such as drones and IoT systems. The research involved data augmentation techniques including crop (0–20% zoom), hue (±20°), saturation (±30%), brightness (±20%), exposure (±15%), and mosaic augmentation. These augmented images were used to train four YOLO nano variants (v5n, v8n, v11n, v12n), which were evaluated using standard metrics: Precision, Recall, F1-Score, and mean Average Precision (mAP). Among the models tested, YOLOv11n with Custom Optuna configuration delivered the highest performance, achieving a 94.6% F1-score and 97.8% [email protected]. These results demonstrate that the optimized YOLOv11n model can support accurate and efficient real-time weed detection in household environments, particularly on edge devices with limited hardware capabilities. This makes it a viable solution for practical implementation in precision agriculture and smart gardening
Performance Comparison of YOLOv8 and DETR in White Blood Cell Detection
No full textAutomated detection and classification of white blood cells (WBCs) from microscopic images play a vital role in supporting the diagnosis of hematological diseases. Accurate and robust object detection algorithms are essential for handling interclass similarities and imbalanced datasets. This study aims to evaluate and compare the performance of two modern object detection algorithms—Detection Transformer (DeTR) and YOLOv8—in performing multiclass WBC classification using public datasets from various sources with diverse visual characteristics. Five experimental scenarios were designed based on varying class distributions and data augmentation techniques, including horizontal/vertical flipping and random rotation. Both methods were trained and evaluated on the same dataset partitions, and their performances were assessed using the following standard metrics: precision, recall, and F1-score for each WBC class. The results show that YOLOv8 consistently achieved superior and more stable performance across all scenarios, with average F1-scores close to 1.00 even in augmented and imbalanced conditions. In contrast, DeTR performed competitively in balanced scenarios but showed lower consistency, particularly in classes such as Neutrophil and Monocyte. Data augmentation positively affected both models, although the gains were more prominent in YOLOv8. This study highlights the strong potential of YOLOv8 in real-time WBC classification tasks and presents DeTR as a viable yet less-optimized approach for this application. These findings contribute to the advancement of medical image-based object detection and offer valuable insights into the selection of appropriate algorithms for hematological image analysi
Explainable Ensemble Learning Framework with SMOTE, SHAP and LIME for Predicting 30-Day Readmission in Diabetic Patients
No full textHospital readmission among diabetic patients poses a significant burden on healthcare systems due to its frequency and associated costs. This study presents a machine learning framework for predicting 30-day readmission in diabetic patients using the Diabetes 130-US Hospitals dataset. The framework integrates data preprocessing, SMOTE for class balancing, ensemble learning, and explainable AI (SHAP and LIME) to enhance both accuracy and interpretability. Multiple models were evaluated, and the best performance was achieved by a weighted ensemble with a recall of 89.43% and an F1-score of 0.6612, indicating strong sensitivity. Explainability analysis using SHAP and LIME highlighted key predictors, notably Medication Change Status and Inpatient Admissions, which are clinically relevant. By combining predictive performance with transparent explanations, the proposed framework offers a practical and trustworthy tool for clinical decision support in managing diabetic readmissions
ResNet50-Driven Quality Inspection for Recorder Musical Instrument
No full textThe manufacturer of a recorder musical instrument requires high-quality product. The aim is to produce precise tones and an aesthetic look at customer satisfaction. A major challenge encountered by manufacturers is traditional visual inspection. Human error is a major factor, notably over extended work periods and the subjective judgment of quality control personnel. This paper reports on the development of a machine vision system for detecting abnormal patterns on the inner surface of a recorder musical instrument. An industrial-grade camera with a resolution of 1280 × 1024, paired with industrial lighting, was utilized. Due to its tube-shaped construction of the object, the bright-field imaging technique is applied to illuminate the interior. ResNet50 was selected as a feature extractor due to its balance between accuracy and efficiency. In addition, a Neural Network served as the classifier. A total of 1,118 images were collected as training data and 304 images as testing data. The training and testing data were separate sets that were taken independently, preventing any risk of data leakage. The test results indicated that the model performed exceptionally well in classification, achieving an accuracy of 95.7%, precision of 95.45%, sensitivity of 96.07%, and specificity of 95.36%. Moreover, the area under the curve of the Receiver Operating Characteristic (ROC AUC) score in test data reached 0.9906, reflecting the model's ability to separate features from the two classes. These findings suggest that the proposed method offers an alternative to subjective visual inspection. Future research may examine diverse deep learning architectures to further enhance performance while achieving faster classification